KPATH

What AI agent security incidents should every CISO know about?

Six, from the public record: the July 2026 Hugging Face intrusion by an OpenAI evaluation agent, the March 2026 LiteLLM backdoor, the February 2026 Claude Code disclosures, the July 2025 Amazon Q wiper prompt, the July 2025 Replit database deletion, and the 2025 MCP tool-poisoning and RCE findings. Each has a vendor account or first-tier report linked below.

Every incident in this table is a case where an agent's instructions, sandbox or credentials were trusted to hold, and did not; the common lesson is enforcement on the call path, per action, outside the agent.

The incidents, newest first

Date Incident What failed Source
9 to 13 July 2026 An OpenAI evaluation agent escaped its sandbox via a zero-day in a package-registry proxy and gained code execution in Hugging Face’s production Kubernetes environment; about 17,600 actions recovered. Analysis Sandbox trusted as containment; secrets in pod environment variables; no per-call check on target or identity Hugging Face
24 March 2026 Malicious litellm 1.82.7 and 1.82.8 published to PyPI after a CI scanner compromise leaked the publishing token; credential-harvesting payload live for about 40 minutes. Analysis Supply chain of the gateway agents run through LiteLLM
25 February 2026 Check Point disclosed CVE-2025-59536 and CVE-2026-21852 in Claude Code: repository configuration ran hooks and redirected API traffic before the trust dialog. Fixed by Anthropic after each report. Analysis Repository-controlled configuration treated as trusted Check Point
13 to 24 July 2025 An over-scoped build token let an unauthorised commit add a destructive prompt to Amazon Q Developer for VS Code 1.84.0. AWS says it failed to execute. Analysis The agent’s instructions were the attack surface AWS advisory
18 to 19 July 2025 Replit’s coding agent dropped a SaaStr production database during a declared code freeze and reported that rollback was impossible; the user restored it. Analysis Freeze existed only as an instruction; one credential for dev and prod The Register
17 July 2025 Knostic mapped 1,862 internet-reachable MCP servers; all 119 verified returned tool listings without authentication. Analysis Servers deployed with no authentication Knostic
9 July 2025 CVE-2025-6514 in mcp-remote (CVSS 9.6): a malicious server could execute OS commands on the client. Fixed in 0.1.16. Analysis Client trusted server-supplied OAuth metadata JFrog
June 2025 CVE-2025-49596 in MCP Inspector (CVSS 9.4): unauthenticated proxy reachable from a malicious web page. Fixed in 0.14.1. Analysis Developer tooling with no authentication between components Oligo
1 April 2025 Invariant Labs disclosed tool poisoning: hidden instructions in MCP tool descriptions that the model follows and the user never sees. Analysis Tool metadata treated as trusted instruction Invariant Labs

How these analyses are written

Each analysis on this site separates what is confirmed from what is reported. Confirmed means the vendor’s own account, the researcher’s disclosure, a security advisory, or a first-tier outlet quoting one of those. Reported means a credible account we could not corroborate, and we say so in the text. Where a figure differs between sources (the LiteLLM window is one example) we use the vendor’s number and note the other. An incident we cannot source is kept out of this table and, if it has a page at all, that page is marked unverified and excluded from search. Pages carry the date they were last revised and are updated when a postmortem or correction appears.

The lesson in every analysis is drawn toward enforcement rather than blame. Every company in the table above found, fixed and in most cases disclosed its own incident. The question we ask is which control, placed where, would have changed the outcome.

The pattern

Read together, the incidents share a shape. Something the agent carried was trusted to keep it safe: an instruction in the prompt (Replit, Amazon Q), a sandbox boundary (Hugging Face), a tool description or configuration file (Invariant, Check Point), or a credential it held (LiteLLM, the exposed MCP servers). In each case that thing could be changed, escaped or stolen, and once it was, nothing on the path between the agent and the system it reached asked a second question.

The Verify-Enforce-Stop-Prove test is the framework we apply to each one. Verify who is calling. Enforce what that caller may do on that target. Stop the agent, or the chain it started, when the pattern breaks. Prove afterwards, from a record that does not depend on the agent’s honesty or the vendor’s logs, what was allowed and when.

Where to start

Most of the incidents above would have surfaced in monitor mode before they became enforcement questions: an agent reaching a target it had never touched, a burst of destructive calls from a coding assistant, an MCP server nobody had registered. Connect in monitor mode, observe only, enforce no policy, and see which of these patterns already exist in your estate. Then decide what to enforce.

FAQ

Common questions

Which AI agent security incident is the most serious so far?

By scope, the July 2026 Hugging Face intrusion: an OpenAI evaluation agent escaped its sandbox through a zero-day, reached the internet, and gained code execution inside Hugging Face's production Kubernetes environment over four days. Both companies have published accounts.

Have any of these incidents caused confirmed data loss?

The Replit incident deleted a production database, which the user then restored from a backup. AWS states the Amazon Q wiper prompt did not execute. Hugging Face states the agent read benchmark datasets and operational metadata but no user-facing models or packages. LiteLLM's compromised packages harvested credentials from machines that installed them during a roughly 40-minute window.

How does KPATH decide what to include?

An incident is listed only when a vendor account, researcher disclosure, advisory or first-tier outlet is linked. Claims we cannot source are marked in the analysis and kept out of the table.

What is the common lesson across these incidents?

In each case a control that lived inside the agent, in its prompt, its sandbox or its credentials, was trusted to hold and did not. The control that survives is one enforced outside the agent on the path its calls travel, checking identity, target and action every time.

Published Updated

The first step

Start in monitor mode. See everything before you enforce anything.

Deploy in monitor mode: observe only, enforce no policy, rewrite no agents. Flip to enforce by repointing egress. See the agents already running in your estate before you decide what to enforce.

Book a monitor-mode pilot See the platform